Nvidia has released software tools designed to improve the safety of AI agents, as technology companies confront incidents involving autonomous systems that can perform complex tasks and potentially access computer systems without adequate controls. The chipmaker said its new tools could have prevented the recently disclosed attack involving Hugging Face, an AI development platform that Nvidia later acquired for $13 billion. The release comes as OpenAI and Anthropic investigate incidents in which their AI agents accessed commercial and government systems. Nvidia’s approach treats agent safety as an engineering challenge, with the company seeking safeguards that can contain potentially harmful behaviour. One tool, OpenShell, uses hardware capabilities in Nvidia’s central processor chips to isolate AI agents and restrict their access to surrounding systems. Nvidia said it is also working with Arm Holdings and Intel to make the technology compatible with their central processors, potentially extending its use across different computing environments. Justin Boitano, Nvidia’s vice president and general manager of enterprise computing, said the platform could have stopped the Hugging Face breach if it had been deployed during early model evaluations at frontier AI laboratories. He said Nvidia was making the technology available openly and wanted the wider industry to participate in its development. Nvidia has also introduced Sentry, a separate security system designed to work with OpenShell. Sentry uses an additional Nvidia chip to monitor agent activity and can cut off a rogue agent if it attempts to escape the container in which it is operating on a central processor. The tools are intended to address a challenge posed by increasingly autonomous AI systems: agents may execute chains of actions, interact with multiple applications and potentially create additional agents to bypass restrictions. Ali Golshan, senior director of AI software at Nvidia, said the company’s technology uses mathematical techniques to identify attempts to circumvent security controls, including situations where an agent could spawn several sub-agents to evade restrictions placed on the original system. The development reflects growing industry attention on agentic AI, in which software systems can independently plan and carry out sequences of tasks rather than simply respond to individual prompts. As these systems become more capable, developers and companies are seeking ways to limit access, monitor behaviour and intervene when they act outside intended boundaries. Nvidia CEO Jensen Huang has opposed broad AI safety regulations, instead comparing the challenge of securing AI systems with the engineering work involved in making automobiles safer. The company’s latest tools represent an effort to address risks through technical infrastructure, hardware-based isolation and monitoring. By releasing the technology and pursuing compatibility with processors from other manufacturers, Nvidia is positioning its safety systems as potential building blocks for organisations developing and evaluating autonomous AI agents.
Nvidia Launches AI Safety Tools After Rise in Agent-Related Incidents
